Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Analytics Group
Authors: Anders Thomsen Rune Sandager Andreas Vig Logerman Jannick Severin Johanson Steffen Haldrup Andersen
Last updated:
Jan 2013
Analytics Group
Table of contents
PREFACE ........................................................................................................................................................................................................ 1 1 INTRODUCTION TO EVIEWS...................................................................................................................................................... 1
1.1 What is Eviews?............................................................................................................................................................................................................................... 1 1.2 Installing Eviews .............................................................................................................................................................................................................................. 1 1.3 The EViews Interface ................................................................................................................................................................................................................... 2 1.3.1 The empty interface ............................................................................................................................................................................................................... 2 1.3.2 Objects and variables in the interface ....................................................................................................................................................................... 2
2
2.1 2.2 2.3
3
3.1 3.2 3.3 3.4 3.5
DESCRIPTIVE STATISTICS...........................................................................................................................................................10
4.1 The Basics ........................................................................................................................................................................................................................................ 10 4.2 One sample t-test two sided ............................................................................................................................................................................................ 13 4.3 One sample t-test one sided ............................................................................................................................................................................................ 15 4.4 Testing for differences in mean based on two groups .................................................................................................................................... 16 4.4.1 False F-Test ................................................................................................................................................................................................................................ 17 4.5 Paired Sample T-tests ............................................................................................................................................................................................................... 21
5.1 The basics ........................................................................................................................................................................................................................................ 25 5.2 The ANOVA test in Eviews ..................................................................................................................................................................................................... 27 5.3 Testing assumptions .................................................................................................................................................................................................................. 29 5.3.1 Homogeneity of variance (1) ........................................................................................................................................................................................ 29 5.3.2 Normally distributed errors............................................................................................................................................................................................... 30 5.3.3 Independent error terms (3) ........................................................................................................................................................................................... 33
6.1 The basics ........................................................................................................................................................................................................................................ 35 6.2 Scatter dot graphs ...................................................................................................................................................................................................................... 36 6.3 Model estimation in Eviews .................................................................................................................................................................................................. 37 6.4 Model output .................................................................................................................................................................................................................................. 39 6.5 Testing SLR assumptions ......................................................................................................................................................................................................... 40 6.5.1 Testing for heteroskedacity SLR.5 ............................................................................................................................................................................ 40 6.5.2 Testing for normally distributed errors ...................................................................................................................................................................... 43
Analytics Group
The basics ........................................................................................................................................................................................................................................ 44 Model estimation in EViews .................................................................................................................................................................................................. 44 Models with interaction terms ............................................................................................................................................................................................. 45 The assumptions of MLR .......................................................................................................................................................................................................... 46 Testing multiple linear restrictions the Wald test ................................................................................................................................................. 47
8
8.1 8.2 8.3
9
9.1 9.2
ENDOGENEITY ................................................................................................................................................................................56
The basics ........................................................................................................................................................................................................................................ 56 IV estimation using EViews.................................................................................................................................................................................................... 57
10
10.1 10.2 10.3 10.4 10.5 10.6 10.7 10.8 10.9 10.10
11
11.1 11.2
12
12.1 12.2 12.3 12.4 12.5 12.6 12.7
13 14
14.1 14.2
14.3 14.4
14.5 14.6 14.7
Analytics Group
Selected Keywords that Return Scalar Values ..................................................................................................................................................................... 97 Selected Keywords that Return Vector or Matrix Objects ....................................................................................................................................... 98 14.8 Equation Methods ................................................................................................................................................................................................................. 98 14.9 Loops ............................................................................................................................................................................................................................................. 99 14.10 Simulation study Monte Carlo Simulation ........................................................................................................................................................ 100
15 16 17
APPENDIX A - VARIABLES IN THE DATASET RUS98.WF1 ................................................................................. 101 APPENDIX B THE DATASET FEMALEPRIVATEWAGE.WF1 ............................................................................ 102 APPENDIX C INSTALLING WINDOWS ON A MAC ............................................................................................ 103
Analytics Group
Preface
Before reading this manual there are a few things you need to be aware of. First of all, this manual is made by the Analytth ics Group (www.asb.dk/AG) to support BSSs 5 semester economic students and Can.merc Finance students in their use of Eviews. It is far from a complete guide on how to use the software, but only meant to support the students with their specific needs. The manual is not a statistics guide or a textbook, and should not be read as a substitute for either. To make this point crystal clear, we will be making references to the two text books set by the professors: Keller. Statistics for management and economics, 8th Edition 2009. Thomson.. Wooldridge. Introductory Econometrics - A Modern Approach. 4th Edition 2009. Thomson. Marno Verbeek A guide to modern Econometrics. 2
nd
The aim of this manual is to show you how you use a specific software application to make statistical analysis. It is not intended to teach you statistical theory. We do believe that to be able to use this kind of software, for making valid and meaningful analysis, one must have a sufficient understanding of the underlying statistical theory. Throughout this manual we will be using eight different work files to illustrate the use of Eviews. The first file is based on a survey made among ASBs students back in 1998. It contains around 20 variables all of which can be found in appendix A at the very end of this manual. The second work-file is from a research on wage differences between the sexes. The variables and their names in this work-file is considered easy to interpret and should not require any further notice. Both of these work files are available on www.asb.dk/AG
1 Introduction to Eviews
1.1 What is Eviews?
E-views is a spreadsheet software used for various types of data analysis. It has some similarity to the commonly used Microsoft Excel and does support this type of files. According to its creators E-views is characterized as: EViews provides sophisticated data analysis, regression, and forecasting tools on Windows based computers". While you are able to conduct some data analysis in Excel, E-views enables you to do traditional Excel analysis, like descriptive statistics, but also more advanced calculations, regressions and simulations, which you wont find in Excel. In addition to its increased fun ctionality, it also operates at a much faster pace, both in terms of calculation time and in terms of ease of use. Especially Eviews data series analysis functions are superior to many of its competitors.
Analytics Group
This manual is based on version 7.0 of Eviews. There might be minor differences from the student version of the application, but these differences will not be touched upon in this manual.
At this point the interface only includes areas of importance: At 1. Is the traditional tool bar, which includes the different tools, used later in this manual. It is important to notice that the content of these different dropdown menus depends on which Eviews window you select beforehand. E.g. not selecting a data set and clicking the proc bottom gives you no options at all, while the same click gives multiple different opportunities after selecting a window containing data. At 2. is the coding area/prompt. This area allows you to apply different text based commands, which is used for both data manipulation and as a potential shortcut for making different regressions. The grey area below the coding line is somewhat similar to the desktop of your PC. It can include numerous windows, including data spreadsheets, regression results, graphs and several different outputs. 1.3.2 Objects and variables in the interface After importing data, making some calculations, graphs etc.(see the following parts of this manual) the interface could look something similar to the following:
Analytics Group
After opening an existing work file, you will see the window at 1. The windows contains a list of all the variables in the work-file, this list is somewhat similar to the columns of an Excel sheet. To view every single observation and its number, one must select the variables of interest by holding down the ctrl bottom and clicking the variables of interest. To the spread sheet window similar to 2, you must either right click the group or click the view bottom, then click open /as group. It is often a good idea to save groups, equations, graphs (called objects) by a specific name. This is done by clicking the name button, which is circled in the picture above. After assigning a name and clicking OK, the object will appear along the variables in the first window. The object will appear with a symbol matching its kind of object (graph, group, equation etc). The order in which you select the different variables, is the same order you get when you open the "group." The above window is achieved, by first pressing variable sp05, then holding ctrl and press sp06 and so forth. When you have selected the 4 different variables, right click, and open as group.
Analytics Group
2 Data Import
Importing data is straight forward as long as the structure of the data file is correct. In general you need to make sure that the data is structured with variable names in the top row of your spreadsheet and then having the observation following below (see the below illustration from Excel)
It should be noted that besides the following ways of importing, Eviews also support several other file types and application for importing, but we will focus on the most common ones.
Analytics Group
Analytics Group
Alternately you may use the coding area command: genr followed by the new variable name and the new variable equation GENR NEW_VARIABLE_NAME VARIABLE EQUATION.
Analytics Group
variable VAR, the function also known as LN() on must calculators. Using the natural exponential function in Eviews is done with the command exp(VAR) similar to using e^x on must calculators. An example using these functions: genr VAR_NEW2 = log(VAR1+1) + exp(VAR2) Note how we add 1 to VAR1 before applying the log function. This is done to ensure that we do not take the logarithm of zero. In general, making operations which are not mathematical possible e.g. dividing by zero or taking LN(0) will result in error pop-up showing, in the middle of the screen.
This use of dummies can freely be combined with the previous shown operations, allowing you to create more advanced resulting variables.
Analytics Group
Also the use of these constraints (=, <,>), combined with the operators AND and OR ,can be used in numerous ways when creating new variables. To illustrate this claim here is a more advanced example, which uses multiple of these constraints. We want a variable which states the height (sp05) of the respondent, but only if its a male , (sp01=2) with a weight of 1 more than 80kg (sp04>80). If not we want the variable to equal the natural logarithm of the height : Genr advanced1 = sp05*(sp01=2)*(sp04>80) +(1- (sp01=2) *(sp04>80))*log(sp05) Note that (1- (sp01=2) *(sp04>80)) will equal 0 if, and only if, both the constraints are true and 1 if not.
Dont try to make any sense of this variable; its nothing but an example of a more advanced way of using the dummy
Analytics Group
Generating this grouping variable can be done in different ways. First we will illustrate how to create this group variable by first creating tree dummy variables (gr1, gr2 and gr3) and next we will show you how to create the same variable without using these dummies. The dummy method is made by using the command line:
genr gr3 = (var >= 4) genr gr2 = (var = 2 or var = 3) genr gr1 = (var <= 1 )
This takes care of the tree dummies. To create the final grouping variable, group, we use these 3 dummies in the following way:
genr group = gr1*1 + gr2*2 + gr3*3
An alternative method is to combine the dummy creation with the above code in one line of code:
genr group = (var>=4)*3 + (var=2 or var=3)*2 + (var<=1)*1
Analytics Group
10
Within this sample range tool, area 2 concerns the observation number of the sample. To include all observation you need to write @All. To include a specific range you simply write the starting point followed by the ending observation in this case the starting point is observation number 200 and the ending point is observation 455. Area 3 allows you to constrain your analysis of the sample by using the previous mentioned constraints. In this case the analysis is reduced to only including observations which has sp02=2, that is all respondents whom expect an annual income above 300.000 Dkr. You can make these constraints more advanced by using the words and & or while adding more constraints. E.g. to only analyse male respondents who expect an annual income above 300.000 Dkr would be: sh01=2 and sh02=2
4 Descriptive Statistics
4.1 The Basics
In the following we use the data set called rus98_eng in use.wf1, which contains information concerning 455 students at ASB, such as grade, age, gender etc. Getting the most basic descriptive statistics in Eviews is very straight forward. First you need to select the variable of interest (i.e. sp09, which is the average grade of the students) by double clicking it, or right clicking and choosing open as group.
Analytics Group
11
Next you need to click View within the new window and select Descriptive statistics & Tests. Doing so gives you a list of different options, as shown below.
Analytics Group
12
That is, the average of sp09 (average grade) is 8,47, and the std. dev. is 0,738. The variance is not directly reported, but can be obtained by: Var(x) = (std dev (x))^2 Var(sp09) = 0,738
2
The number of observations is 445. Be aware that this means that 10 of our original respondents have not answered the question sex. Histogram and stats reports similar stats, but includes a distribution histogram:
The stats by classification will report statistics grouped by another variable, sp01, sex in this case (sex equals 1 if it s a female, and equals 2 if its a male):
Analytics Group
13
The output above shows that in our sample the 170 females have a mean average grade of 8,53, while the 275 males in the sample have a mean average grade of 8,44.
Analytics Group
14
To make this test in Eviews we first select the variable called sp09 (average grade) by double clicking it, and then choose
Then we have to type in the value from H0. In our example our H0 hypothesis is that the average grade equals 7. Thus we write the following, and click OK.
Analytics Group
15
This gives us our t-statistic on 42,19, which we have to compare to the critical value of the t-distribution with 444 degrees of freedom at a significance level of 5%. Because we are applying a two-sided test, the critical values are -1,96 and 1,96. Our conclusion is that we reject our null hypothesis because the test statistic falls in the critical region. The very low p-value indicates that our conclusion is not sensitive to changes in the significance level.
Analytics Group
16
Our test-value is 8,3, so in Eviews we simply write: 8.3 (We use . instead of , )
This gives us a t-statistic of 5,03, which we have to compare to the critical value of the t-distribution with 444 degrees of freedom and a significance level of 5%. Note that this time we apply a one-sided test and the critical value is changed to 1,645. Thus we reject the null hypothesis. Our conclusion is not sensitive to changes in the significance level. Note that the p-value in Eviews refers to a two-sided test!
H0: grade_men= grade_women grade_men - grade_women = 0 H1: grade_men grade_women grade_men - grade_women 0
Analytics Group
17
We want to test if the average grade for females is different from the average grade for males. The first step is to determine if the variances are equal. To determine this we could either use the False F-test or Levenes test. (To test for equality in variance using Levenes test please see the section about ANOVA).
4.4.1 False F-Test To conduct the false F-test, Eviews is used for calculating the sample variance for each group. This is done as shown in section 4 descriptive statistics. Eviews does not contain the test by default. The hypothesis looks like this: H0: 1 = 2 = ... = K
2 2 2 2
We then want to know if the variance in variable sp09 (average grade) is the same for the two groups (male, females). The False F-test looks like this if we only have 2 groups:
We then need to find the largest ( ) and smallest ( sp09 (average grade) and do the following:
Analytics Group
18
Our group variable was sex (sp01) The resulting output contains the mean, standard deviation and number of observations by group (Recall that = Variance):
2
Then we have to calculate our F-statistic by dividing the largest sample variance (0,753 = 0,567) with the smallest sample variance (0,712 = 0,507), and compare this to a critical value:
2
Analytics Group
19
The critical value of the F-distribution for a two sided test with 169 and 274 degrees of freedom at a significance level of 5% can be found to be: = 1,31. This indicates that we cannot reject the null-hypothesis, and therefore we conclude that the variance is equal across the two groups.
If you have more than two groups, the critical value is calculated a bit different. The test is still level has to be adjusted. We use
( )
Let's say we have 5 groups instead of only 2. If this is the case, the critical value of the f-distribution can be found as: where And therefore: = 1,47. Now let's continue with our test for the difference in means based on two groups. If you need to test more than two groups, you need to use another test, such as ANOVA. Given equality of the variance, the second step is to determine if the means are equal. The test is performed in Eviews by choosing the variable of interest (sp09 = grade) and then view Descriptive Statistics and Tests Equality Test by classification =
( ) ( )
Analytics Group
20
Our grouping-variable in this example is sex (sp01). We want to test the null hypothesis, that the 2 means are equal, which is why we choose mean in the window:
This gives us a t-statistic of 1,30, which we have to compare to the two-sided critical values in the t-distribution with 443 degrees of freedom, and a significance level of 5%. This critical values can be found to be approximately equal to -1,96 and 1,96. This indicates that we cannot reject the H0 hypothesis. Therefore we conclude that there is no difference between the average grade of males and females. The p-value is relative high, which indicates that the conclusion is not sensitive to changes in the significance level.
Analytics Group
21
Analytics Group
22
These output tells us that production_b has a larger sample mean than production_a. When we have a paired sample ttest, the hypothesis looks as follows:
H0: produktion_a = produktion_b produktion_a - produktion_b = 0 H1: produktion_a produktion_b produktion_a - produktion_b 0
Analytics Group
23
First of all we have to construct a new variable, which measures the difference between the two daily productions. In Eviews we make this variable in quick Generate Series:
First we have to name our new variable. In this example we just name it Difference. We then have to explain to Eviews how this new variable should be calculated, so we write: Difference=production_a-production_b
The new variable is constructed, and now we can use the simple hypothesis test to do paired sample t-test. Start by opening the new variable Difference and then choose view Descriptive statistics and Tests Simple hypothesis test
Analytics Group
24
Please recall our problem. We were interested in testing the hypothesis if there was a difference between the two different harvesters. Our test-value in our example will therefore be 0, just like the hypothesis.
Analytics Group
25
This gives us a t-value of -0,46, which we have to compare to a critical value of t-distribution with 11 degrees of freedom at a significance level of 5% for a two sided test. This critical value is +/-2.228, which tells us that we cannot reject the null hypothesis. Thus, we cannot conclude that there is a difference between the two harvesters. The very high p-value indicates that our conclusion is not sensitive to changes in the significance level.
H0: 1 = 2 = = k
Against the alternative
Analytics Group
26
The sample mean and standard deviation is calculated, like shown in the descriptive statistics section, by selecting the variable by double clicking it and going View/ Descriptive Statistics/Tests/Stats by classification.. In the resulting window, the parameters of interest, mean and standard deviation in this case, and name the grouping variable political party, sp03.
Analytics Group
27
The resulting output contains the mean, standard deviation and number of observations by group:
To determine if there is a statistically significant difference, we need to run an ANOVA test, which is shown in the following section.
Analytics Group
28
View > Descriptive Statistics & Tests> Equality Tests by Classification... (see the picture below)
To let Eviews know that we want to group the variable based on the political party, which is variable sp03, we type in sp03 as shown above. The Test quality of is set to mean as default, so we simply leave this setting. Clicking OK will give us the desired ANOVA output:
Analytics Group
29
To determine whether to reject the null hypothesis or not we focus on the highlighted ANOVA F-test output. The column named Probability contains the p-value of interest. Since the p-value is below 5% we reject the null hypothesis and conclude that there is a statistical significant difference in weight between the groups.
To have EViews run Levenes test, is somewhat similar to running the ANOVA test in the first place. Once again you need to select the variable of interest, sp04, and then go:
Analytics Group
30
In the resulting window you once again put in the grouping variable, sp03, but this time you ask Eviews to Test equality of Variance and not mean.
Just like in the ANOVA case we base our conclusion on the resulting p-value. But unlike in the ANOVA case we get a pvalue of 0.67, which is way above any reasonable level of significance. Therefore we cannot reject the null hypothesis and assumption of homogeneity of variance is considered satisfied. To test for the homogeneity of variance you could also use the False F-test. This is done in section 4.4.1 .
5.3.2 Normally distributed errors We address the issue of normality within each group. Doing so can be done in different ways. First we address the assumption by creating distribution histograms for each group. Doing so is done by first selecting the dependent variable, the weight, sp04, by double clicking it. Then clicking Graph will result in the following option window:
Analytics Group
31
To make Eviews group the observation, first select Categorical Graph which gives you the additional option to the right. Then select Distribution and make Eviews do the actual grouping by writing the variable name sp03, in the Across graphs window. The result should look similar to this:
An alternative way of checking for normality is doing so across the different groups. Making this cross group analysis is done by using Q-Q plots to determine whether or not the observations follow a normal distribution when analyzed within their group. To make this analysis in EViews do the following:
Analytics Group
32
Select Quick > Graph from the top tool, which should result in the following windows:
Within this windows type in the variable of interest, the weight sp04 (in this case), and click OK and you will face the following option window:
First you need to choose Categorical Graph from the dropdown menu (1). Then select the specific graph Quantile Quantile (2), which is also known as the Q-Q plot. To make Eviews create a separate graph for each outcome in the grouping variable, you need to type in the grouping variable in the Across Graphs window. If you for some reason want Eviews to test for another distribution then the normal distribution you can change the options of the test in the Details window, but this is not of interest in our example. After clicking OK the resulting output presents itself:
Analytics Group
33
The output displays the perfect normal distribution, the red line, and the actual observations, the blue dots, within each group. As we can see there exists only minor deviations from the red line and therefore we conclude that the assumption concerning normal distributed errors is satisfied. Before rejecting the assumption of normality one should always consider the properties of the central limit theorem see Keller p. 300 for further. 5.3.3 Independent error terms (3) Problems concerning this assumption is by construction rarely a problem when analysing cross-sectional data, but is still mentioned in this manual to illustrate how the assumption is treaded in Eviews. Assumptions concerning independent error terms is simply done, by making scatter plots of the variable of interest and the observation numbers. This is done to ensure that a pattern related to order in which the sample is collected, doesnt exist. Making a scatter plot diagram like this is somewhat similar to the graphs made above: Select Quick/Graph and type in the variable of interest in the resulting window and click OK.
Analytics Group
34
This time you need to leave the option at Basic Graph and select Dot Plot from the specific window. Before clicking OK make sure that you window match the one shown in the picture above.
Analytics Group
35
The resulting output is shown above. Since there is no evidence of a pattern cross the observation numbers, we conclude that the assumption concerning independent error terms is satisfied. Once again note that Eviews doesnt report the errors and we used the actual observations. In this case this using the actual observations should not make any difference.
1) Is
2) What is the magnitude of In this and the following section we will be using the work-file FEMALEPRIVATWAGE.wf1 to determine the relation b etween a persons hourly wage and the co-variants such as education, experience, marriage and children. To illustrate how SLR could be used in this framework, consider the following example: We are interested in determine how education ( educatio in the work-file) is related to hourly wage (hourwage) and thus the relationship:
Analytics Group
36
That is determining how large, if any, an effect education has on the hourly wage.
Please note:
The implications, theory and challenges concerning simple linear regression analysis are the main topics of Wooldridge chap. 2. In the following we will assume that the content of this chapter is somewhat known theory and we will not go into detail with more advanced implications of SLR such as reverse causality or omitted variable bias.
Here you simply type in the names of the two variables of interest, the order is not important, and click OK.
In graph type, select the default Basic Graph and Scatter from the specific list. To make sure that your variables are on the right axis, click the Axis/Scale tab and adjust the axis like shown below.
E.g. lets say you plot y against both x, ln(x), x^2 and discover a better fitting relationship.
Analytics Group
37
The resulting output should look somewhat similar to the one shown here. Note how we made the scatter plot for both hourly wage and LN(hourly wage) against education. This is done to show how scatter plots can be used to explore different relationships between variables. In this case we find that the LN(wage) vs. education show evidence of a better fitting relationship than wage vs. education . Based on this observation we could consider rewriting our model to the form: ( )
Analytics Group
38
The Equation Estimation tool is shown in the previous picture. Most importantly is the Equation specification in which you have to specify which variables you want to include and their internal relations. It is really important to name your dependent variable (often called y) first. Following your dependent variable you should type the constant ( ) written in EViews as c for constant (forgetting this constant has huge implications on the resulting estimation). Your regre s6 sior/explanatory/independent variable (also known as x) should follow the constant c (the from our model) . In the section marked as 2 in the above picture is where you tell EViews which statistical method it should use to estimate the equation. In this manual we will not cover other than the default setting called least squares (see Wooldridge p. 27 for how to Derive the Ordinary Least Squares - the use of more advanced methods such as maximum likelyhood and two state least square is covered in Wooldridge). Clicking the options tab (marked as 3) results in the following equation option window:
Analytics Group
39
Especially the marked area concerning heteroskedasticity might be of relevance somewhere down the line, since its used to adjust for the problems concerning heteroskedasticity (Wooldridge chap. 8). When the equation has been specified and all options are in place, clicking OK will result in an output similar to the one shown and described in the following section.
Analytics Group
40
The output produced by Eviews (see the above picture) might seem overwhelming at first, but far from all of the data has 7 relevance at this stage . We have highlighted the most important aspects of the output to make the interpretation easier accessible: 1. 2. Section shows the dependent variable (your y) always make sure this part of the output is as intended. The output reports that LOG(HOURWAGE) is the dependent variable just as intended. Section contains the most critical information. The table shows the estimated constant, , and the estimated coefficient, , their standard deviations, t statistics and resulting p-values. In this case the output reports that: Resulting in a final model estimate: ( 3. 4. 5. )
(0,10905) (0,009343) Shows R-squared and R-squared adjusted. Both of which are so call goodness-of fit indicators. To understand the difference between the two see Wooldridge p. 199-203. S.E. regression is useful when the model is used for forecasting/prediction see. Wooldridge p. 206-9 on Predic-
7 8
In this case stage is not only a reference to the indended user but also the SLR. In relation to validity the most importance of the assumptions is SLR.4. that is ( ) see Wool. Chap 2.
Analytics Group
41
The window shows a list of possible tests, all testing for heteroskedasticity. The tests covered in Wooldridge are the Breusch-Pagan-Godfrey [Wooldridge p. 273] and White [Wooldridge p. 274]. No matter which test we use for testing heteroskedacity, the null hypothesis is identical: ( ) Theres homoskedacity
Analytics Group
42
- Theres heteroskedacity
What both tests does is using the squared residuals (RESID^2) as the dependent variable and try to determine whether these can be explained using different forms of the original independent variables [see. Wooldridge chap. 8 for further detail]. To conclude whether we have to reject the null hypothesis or not, using the resulting F statistic is enough. The F-test tests for the joint significant of all the included independent variables (see the future sections on this topic and Wooldridge chap. 4). If these are not jointly significant, then we cannot reject the null hypothesis and assume homoskedadicity. To reject the null hypothesis we would need a prob. value (or p-value) less than 0.05. None of the two tests reports p-values anywhere close to 5% so cannot reject the null hypothesis in other words, heteroskedacity does not seem to be a problem.
Analytics Group
43
6.5.2 Testing for normally distributed errors To test for normal distributed errors we use the Jarque-Bara test for normality. The hypothesis of the Jarque-Bera test is a follows:
Running the test in Eviews is somewhat similar to running the tests for heteroskedacity. First you must estimate the model (which creates the residuals on which the test is based), then simply go: View/Residual Tests/Histogram /Normality test
To determine whether the assumption of normal distributed errors are satisfied or not, we once again turn our attention to the highlighted test statistic and p-value. The p-value in this case turns out to be 0, and as a result, we reject the null hypothesis
Analytics Group
44
The purpose of including more than one variable, to explain the variance in y, are as follow: 1) Deal with the problems of omitted variable bias that is to make the everything else equal assumption SLR.4 valid see. Wooldridge Section 3.1 on motivation for multiple regression. 2) Including more variables, and thereby increase precision of forecasting. i.e. the explanation power of the model, which can increase the
Wooldridge dedicates a large part of his book to this subject. To gain a basic understanding of the concept reading chapter 3 will get you started. But to gain a sufficient understanding reading chapter 3-9 is strongly recommended.
Analytics Group
45
We are still trying to determine why some people make more money than others (the variance in hourwage), but this time we include more potential explanations (independent variables). Note that married and child are both dummy. Married is 1 if a person is married and 0 otherwise. Child is 1 if a person has at least one child has at least one child and 0 otherwise. Running the above equation results in the following output:
When running MLR model estimation the first place to look in the output, is at the F-statistic and its p-value (underlined in the above figure). As described in Wool. Section 4.5 Testing Multiple Linear Restriction: The F Test the F-test tests multiple linear restrictions. In Eviews the hypothesis tested by the F-test in the basic MLR estimation output is:
A simple interpretation of the null hypothesis is that the union of all used regressors do not have a significant effect on y. In our example we find that the regressors used to have a significant effect on y (the p-value is 0 thus we reject the null hypothesis). The analysis of each specific variable, their significance and effect on y is somewhat similar to the analyse SLR see the above section.
Analytics Group
46
year of education has a different effect on hourly wage when employed in the province vs. working in the capital. To determine whether or not this is the case, we need to estimate the following model: ( ) ( )
If is significant different from zero we must conclude that the effect of education does differ depending on your location of work. To run this form of model in Eviews we can either construct a new variable, like shown in a previous section, and then run the model estimation or we can do the following: In the estimate equation window type: log(hourwage) c education province educatio*province Thus giving the output:
Thus we conclude that the effect of education does not differ depending on whether one lives in a province or not.
Analytics Group
47
We suspect that almost all of these variables are somewhat positively correlated with each other. To test for joint significance one option would be to use the F-test as described in Wooldridge p. 143, another option would be to use the Wald test in Eviews. The Wald test tests one or more linear restriction on the model. Let say we want to test the join significance of age and experience in the above example (note that just because one variable is significant does not necessary mean that the group including the variable is significant). The hypothesis in that case would be:
To run the Wald test in Eviews is done by first estimating the model including all reggressors of interest (see the former section on estimation), within the resulting output window go: View > Coefficient Tests >Wald - Coefficient Restrictions like shown in the picture below.
The resulting window is where the restrictions from our hypothesis are written. Doing so can be a little misleading since Eviews names the variables a little different than what is normally done. Eviews names the variables according to the number by which they appear in the output (or in the estimated equation for that matter). So the constant (called C in Eviews) is in this case not C(0) but C(1) since it appear as the first variable on the list.
Analytics Group
48
Writing the restrictions is done in the white field like shown above. Note that we write more than one restriction by separating each with a comma like in the above example. An alternative to the above used way of writing our restriction would be to write: C(5)=C(4)=0 The resulting output when running the test (clicking OK) is shown below:
Eviews reports both the Chi-square and the F-statistic statistics. The choice between the two should not make that big of a difference, since the resulting p-value will not differ by any significant amount. Like in any other hypothesis test we reject the null hypothesis if the resulting p-value falls below our predetermined level of significance (we use 0.05). In this specific case we get a resulting p-value equalling zero and thus reject the null hypothesis and conclude that the variables does have a jointly significant effect on our dependent variable.
Analytics Group
49
The above equation states that the current value Yt equals to a constant plus times its previous value plus an unpredictable component . The equation above is called an auto-regressive process of order p, or in short, AR(p). In this case we are interested in determining the best ARMA representation of D, which is defined as dividends. This gives us the following equation:
If the null hypothesis H0: =0 (or the first autocorrelation = 1) is true, then a unit root is obtained, which indicates that the time series is non-stationary. To test the null hypothesis that =1, it is possible to use the standard t-statistic, but with different critical values calculated using Dickey-Fuller. The DF test is estimated by using three different equations, as presented in E-Views. The three test equations are:
The first equation has an intercept, indicated by the parameter , represents a random walk model with drift. The second equation with a trend ( ) and an intercept represents a random walk model with drift around a stochastic trend, and the last one represents a random walk model.
Analytics Group
50
The data analysis for testing the unit root can be done as follows. Click the div variable, and the values will appear on the screen. Click View/Unit root test and the options for the unit root test is as follows:
Figure 1
Test type:
Different types of statistical options are available. The most common tests are the Augmented Dickey-Fuller, Philips-Perron and Kwiatkowski-Philips-Schmidt-Shin (KPSS). Level, 1 difference and 2 difference. These options are related to the amount of times that the series have to be differenced before it gets stationary. The hypothesis goes as follows: Level: st 1 : nd 2 : I(1) vs. I(0) for H0 vs. H1 I(2) vs. I(1) for H0 vs. H1 I(3) vs. I(2) for H0 vs. H1
st nd
In general, an I(d) process is a series that is stationary after differencing d times. Lag length: According to Wooldrigde J. M. the number of lags included is a trade-off between losing power and a wrong test statistic. There are no general rules to follow in any case, but for annual data, one or two lags usually suffice. For monthly data, we might include 12 lags.
To choose at which level the test should be executed, an initial visual inspection is required.
st nd
Figure 2 illustrates div displayed in level, 1 difference and 2 difference. The visual inspection tells us that the dividends st are trended, either by a random walk with drift or a deterministic trend. Furthermore it seems stationary by taking the 1 difference.
Analytics Group
51
Figure 2 By choosing the following two setups, we intend to do a test on the two auxiliary equations below, which represents the equation with zero lag (standard Dickey-Fller) and one lag respectively (Augmented Dickey-Fuller). We could of course extend the test including more lags, but remember the above-mentioned conclusion.
The Durbin-Watson statistics is close to two, including one lag, which indicates that the errors are serially uncorrelated against the alternative that they follow a first order autoregressive process. By choosing different test types according to Figure 1, we can increase the robustness of the test. As mentioned above, the Phillips-Perron and the KPSS test is preferred. The hypothesis for the ADF and the Phillips-Perron is the same, which means that the null hypothesis claims that a unit root is present, i.e. the more negative the test statistic is, the stronger is the probability of rejecting the null hypothesis; that there is a unit root at the given level of confidence. The KPSS test works the other way around, which means that it tests the null hypothesis of stationarity against the alternative of a unit root. The first table test whether div is I(2) or I(1) (i.e. first difference with intercept), and the second table test whether div is I(1) or I(0) (i.e. in levels with trend and intercept).
Analytics Group
52
The test statistics is all above the critical limit, therefore, we reject I(2) in favor of I(1). *The limits viewed in E-views is not correct, instead selected percentiles of the appropriate distribution are developed and published in several works by Dickey and Fuller. Test ADF Philips-Perron KPSS Test statistics -3,497 -2,931 0,174 5 % critical limit* -3.45 -3.45 0,146 Conclusion Reject Fail to reject Fail to reject
The unit root tests in levels show some different results. The Philips-Perron and KPSS tests suggest a unit root at the 5 % significance level, while the ADF test rejects a unit. As we have two tests pointing I(1) and the ADF test is only marginally rejecting a unit root, it could be sign of dividends having a unit root in levels.
Or even simpler:
This corresponds to the most simple first-order autoregressive growth model. It says that the current value Yt equals to a constant plus times its previous value plus an unpredictable component . The first equation above is called an autoregressive process of order p, or in short, AR( p), and the second equation above is called an auto-regressive process of order 1, or in short, AR(1). To estimate the equation above, which actually could be specified as y c y(-1), either by Quick -> Estimate Equation and then type the equation or type ls y c y(-1) through the command field. But this option do not allow for the time-series part of Eviews. Instead one can use the following possibilities:
y c ar(1), y c ar(1) ar(2) y c ar(1) ar(2) ar(p) y c ma(1), y c ma(1) ma(2) y c ma(1) ma(2) ma(p) y c ar(1) ma(1), y c ar(1) ar(2) ma(1), y c ar(1) ar(1) ma(1) ma(2) y c ar(1) ar(p) ma(1) ma(q)
Lets take the previous discussed data into account, and estimate the best ARMA(p,q) representation of dividends (Dt). To determine the best representation, its important to regress our model upon a stationary process, and as proven we should use instead.
Analytics Group
53
To estimate the equation equation type ls d(div) c ar(1) in the command field. The d in front of div is a Eviews command for taking the first difference. The second difference is easily done by adding another d, as in d(d(div)). The obtained statistical results are as follows:
Figure 3
1. 2.
3.
We fail to reject the null hypothesis of no first-order autocorrelation, H0: , with a p-value of >5%. The point estimator is with std. Error 0.088. The DW statistic is 1.92, which indicates that the model only has a little problem towards a positive autocorrelation in the residuals. This can be tested further by a Breush-Godfrey serial correlation LM test. a. In the interpretation window click view/Residual Diagnostics/Serial Correlation LM test b. Enter 1 lag, for testing H0: (no AR(1) in the error terms). The number of lag is as previous discussed a trade-off, but because of the fact that the data is annual, we are using one lag. c. The p-value is p=0,0174 indicating first-order serial correlation of order 1. The statistical results in the figure below can be obtained by selecting View/Representations. This figure shows the estimation command and equation, as well as the regression function.
4. The next step is to repeat the previous steps, and through trial and error finding the best representation of the dependent variable, in this case dividends. 5. As a help, or an indicator, one can use the correlogram Q-statistics as shown below. Furthermore some residual tests are appropriate through view/Residual Diagnostics.
Analytics Group
54
Correlogram Q statistic The figures 4 - 7 shows three statistics. 10.6 10.6 10.6 The AC (autocorrelation coefficient) The PAC (partial autocorrelation coefficient) A box-pierce Q-statistic with its probability
The lines in the graph of AC and PAC approximate two standard error bounds. Figure 4 represents an AR(1) model with Durbin-Watson statistic 1.923. The graph shows that at lagged k=2, the hypothesis of no autocorrelation is rejected. The Qstatistic is a test statistics for the joint hypothesis that all of the autocorrelation coefficients up to certain lagged values are simultaneously equal to zero. The results in figure 4 show that is rejected up to and including 4 lags. If the mean equation is correctly specified, all Q-statistics should not be significant. However, there remains the practical problem of choosing the order of the lagged variables to be utilized for the test. As you can see the model in figure 5, AR(2), has a Durbin-Watson statistic close to 2, which is recommended, and according to the correlogram the model seems correctly specified. The ARMA(2,1) model in figure 6 does have some problems similar to the discussed above. Figure 7, the ARMA(2,2) could also be correctly specified, but the model offers more lags than necessary.
Analytics Group
55
Analytics Group
56
9 Endogeneity
9.1 The basics
) If one of our explanatory variables is correlated with our error term the assumption of ( does not hold. The problem of endogeneity causes our OLS estimator to be inconsistent and biased. Endogeneity can be present in models with measurement error, missing variables, and or simultaneity/reverse causality. To solve this problem we need another estimator which is consistent and unbiased even though we have endogeneity. The problem of endogeneity can be solved by using instrumental variables. Instruments are variables that are uncorrelated with the error term, and correlated with the endogenous variable. Consider the model without endogeneity
With the moment conditions ( ( Isolating the error term and inserting yields {( ) } {( ) } by replacing the expectation with the sample moments. ( ( ) ) ) )
and
) But these moment conditions does not hold anymore because of endogeneity, namely ( is violated. The solution is to find an instrument (or vector of instruments), to replace the endogenous variable above. The error term and the instrument should be uncorrelated and the instrument and the instrumental variable should be correlated. If is the endogenous variable we have ( ( ) ) , validity criterion , relevance criterion
Analytics Group
57
It should now be possible to solve for the consistent estimate of and , and thus we have the IV-estimator instead of the OLS-estimator. Our IV-estimator can generally be written (in matrix notation) as ( )
Where Z is an NxK matrix with ith row . It will only be possible to use the IV-estimator in the exactly identified case, where number of moment conditions (R) equals number of parameters to be estimated (K) . If we have more valid and relevant instruments, it will be inefficient to throw some of them away and thereby only use a subset. To solve this problem we use the two-step least square (2SLS) estimator. The 2SLS estimator is an IV estimator. In a just-identified model it simplifies to the IV estimator given above. The 2SLS estimator is given by [ ( ) ] [ ( ) ]
The 2SLS estimator gets its name from the result that it can be obtained by two consecutive OLS regressions: OLS regression of x on z to get followed by OLS of y on which provides .
Estimating the wage equation by OLS (Quick/Estimate equation) (potentially inconsistent and biased), we get the following output
Analytics Group
58
We suspect that we have endogenous variable educ. To check the relevance of a potential instrument we make an auxiliary regression. By regressing educ on all the explanatory variables and the potential instrument nearc4, we can examine the relevance of the instrument nearc4. nearc4 is a dummy variable equal to one if the individual lived close to a college, and is therefore assumed to be positively correlated with educ but uncorrelated with the error term, making it a good candidate for instrument. The auxiliary regression is performed by regressing educ with a constant and all explanatory variables, including the potential instrument.
We see that the coefficient on nearc4 is significant, making it a relevant instrument. Well now estimate the regression equation by 2SLS, using the exogenous explanatory variables as instruments for themselves and nearc4 as instrument for educ. In Quick/Estimate equation the following is entered:
Analytics Group
59
Standard errors for 2SLS are bigger than for OLS. It could indicate a weak instruments problem. And our coefficient estimates differ in the OLS and the 2SLS case, which indicates a problem with endogeneity, since they both should be consistent without endogeneity. Testing the exogeneity (making it a valid instrument) of the instrument educ by clicking View/IV Diagnostic & Tests/Regressor Endogeneity Test, type the endogenous variable educ and click OK.
Analytics Group
60
Where claims that there are no differences between the model in which educ is treated as endogenous and the model where it is treated as exogenous. We cannot reject and therefore educ can be considered exogenous. We have options in checking the validity of instruments if we have an over identified model. To construct an over identified model well use meduc as additional instrument for educ. Estimation results:
Again we start by testing for exogeneity, by clicking View/IV Diagnostic & Tests/Regressor Endogeneity Test
There is no clear answer to this test. We cannot say by any certainty that our educ can be treated as exogenous. In our over identified model we will now check for validity of instruments using the C-orthogonality test, found under IV Diagnostics & Tests, and we need to test the instrument one by one. Starting with meduc we get:
Analytics Group
61
We get the same J-statistic in both tests; this is a general result when we only have two instruments and one endogenous variable. Both tests strongly indicate validity of our instruments. Another way of testing the validity of the instruments is the LM-test. To calculate the LM-statistic we need to run an auxiliary regression with the residuals from the 2SLS estimation as dependent variable. To save the residuals from our 2SLS estimation we press Proc/Make residual series, type a name for the residuals and click ok. We regress the saved residuals with regards to all explanatory variables and instruments (excluding the endogenous variable). This provides the following output:
Analytics Group
62
Analytics Group
63
Where ( ), In this represents the covariance matrix, the coefficient matrix, vector containing the intercepts, is a vector containing all our error terms.
is the
Before making the VAR model we should check if our variables are stationary, and thereby not having a stochastic or deterministic trend, so we avoid spurious regression. Open up the workfile. Mark all the variables you want to include in your VAR model. Right click and press Open/as VAR. The following dialogue box will appear:
Analytics Group
64
It will be possible to estimate both an Unrestricted VAR and a Vector Error Correction model. In this model we choose to include two lags of our dependent variable. We could also include more lag intervals. If we instead had written 1 2 5 6 in lag intervals, we would include lag 1-2 and 5-6. The determination of lag length is a trade-off between the curse of dimensionality and reduced models, which are not appropriate to indicate the dynamic adjustment. If the lag length is to short, autocorrelation of the error terms could lead to apparently significant and inefficient estimators. Therefore, one would receive wrong results. The results we get by estimating the model is (with lag interval 1 1, and sample set to 1901 to 2000):
9
Even with inclusion of a small lag length interval we will have to estimate many parameters. Increasing number of parameters causes the degrees of freedom to decrease.
Analytics Group
65
The coefficient estimate is presented without brackets, the number in brackets is the p-value from the significance test of the parameter estimate, and the number in square brackets is the critical value. In this VAR model the data generating process for R is estimated to
Note: none of our variables are significantly different from zero at a 5% significance level in the above example. In the bottom of the estimation output we have the Akaike and Schwartz information criterion. One way to find the right specification and the right number of lags in our VAR model, is to minimize the information criteria.
10.3 Stationary
To check if our VAR(1) model is stationary, we choose the option View/Lag structure/AR root graph
Analytics Group
66
All inverse roots smaller than 1 indicates that our VAR model is stationary.
test.
Where our
is that we do not have Granger-causaility. In our sample we get that DP causes R, and DP causes DD
Analytics Group
67
It is possible to specify which variables we want to give an impulse, and which variables we want to see the response from. In our model, we will get a combined 3x3 graph with the above selection.
Analytics Group
68
The first graph shows how the variable R react to a shock to the variables R, DD, DP. The response of a shock to R is a big positive change in R, the following effect the first period the response is negative. For a shock to DP the response of R is positive, and the effect dies out over time, meaning that a positive shock to log dividend-price ratio will cause a positive shock to the log stock return.
10.6 Forecasting
There are different forms of forecasting. In-sample forecasting uses a sample within another sample which means, that we only use a subset of the data. Then we use the rest of the data to compare our forecasts to the real data. This method is used if we want to evaluate our forecast.
Analytics Group
69
Out-of-sample forecasting uses all the data from the sample, and forecasts in periods ahead of the sample, this is considered real forecasting. We would like to evaluate our VAR-models ability to forecast, therefore we use in-sample forecasting. Choose proc/make
model/solve
It is possible to use two different types of simulation dynamics, dynamic solution and static solution. Static dynamics forecasts one period and updates the information in the forecast, and thereby uses a rolling window of data. Dynamic sol ution constructs an h-period forecast based on data in the sample, and thus does not update the information after each forecast.
Solution sample should be set to the time period we want to forecast. For this example we use 1998 2008. By clicking OK, we will get 3 new variables named after the original variable, but now we an added _0 after. It is possible to rename the new variables by double clicking on them, and choosing Name in the menu, by doing that we secure that these forecasts are not being overwritten by new forecasts. It is therefore possible to make two new variables for every variable that we want to forecast. One variable contains the static forecasts and one containing the dynamic forecasts. Graphing these forecasts is done by clicking Quick/Graph
Analytics Group
70
Entering the variables we want to graph in the above window, where the dynamic forecast for r has been renamed to r_0_dyn. Clicking OK two times yields
The static and dynamic forecasts get very different results as shown above. The dynamic forecast converges to the conditional mean in the long run.
Analytics Group
71
The determination of lag length is a trade-off between the curse of dimensionality appropriate to indicate the dynamic adjustment.
10
If the lag length is too short, autocorrelation of the error terms could lead to apparently significant and inefficient estimators. Therefore, one would receive wrong results. One way to examine if the lag length is correctly specified: View/Lag structure/Lag length Criteria The appropriate amount of lags to be included could be 5 or 10, including more lags causes the loss of more data, and will change the information criteria. We would like to minimize the information criteria. The information criteria function are functions of the log-likelihood function (the more negative the better) and some positive addition that tries to control for the inclusion of more variables. Thereby the information criteria seeks to handle the tradeoff between a parsimonious model and a comprehensive model.
The SC and HQ information criterion chooses 1 lag, and LR, FPE and AIC chooses to include 3 lags. If we only include one lag, we can check the correlogram for autocorrelation.
10
Even with inclusion of a small lag length interval we will have to estimate many parameters. Increasing number of parameters causes the degrees of freedom to decrease.
Analytics Group
72
We test autocorrelation with 5 % confidence interval. Our goal is to eliminate all autocorrelation. We only have three autocorrelation values outside of our confidence intervals and therefore we have sufficiently dealt with the problem of autocorrelation.
Analytics Group
73
In this VAR framework we are able to test for cointegration in a way that does not have the shortcomings of the EngleGranger approach. In the Johansen cointegration test the result does not depend on which variable we normalize with regards to, and it is possible for us include to more cointegration relationships. In the Johansen cointegration test we exploit that the number of non-zero eigenvalues is at most the rank of the matrix , meaning that we can interpret the number of significant eigenvalues as the number of cointegration relations. For the likelihood ratio test we can exploit that the maximum log-likelihood can be expressed as a function of the eigenvalues of . For this example we use a new data set with 3 non-stationary I(1) variables. Variable
Description Annualized interest rate on 3-month treasury bills Annualized interest rate on 1-year treasury bills Annualized interest rate on 10-year treasury bills
The dataset is called tbills.wf1. We would like to find a linear relationship between our variables that is I(0) and thus s tationary. Finding a linear relationship that is I(0) can solve the problem of spurious regression. We should carry out the Johansen test with the correct number of lags to eliminate serial correlation. As showed earlier we could check the correlogram of the functions, to see if we have eliminated the serial-correlation. To carry out the test for cointegration we for instance first estimate our VAR(p) (in this example p=1) model in the usual way, by marking the variables, right clicking and selecting open/as VAR. In this example we choose the lag interval to be 1 1, by clicking OK we get our VAR(1) model, and the vector autoregressive estimates. Before making the Johansen test we should always make a test to see if our variables are I(1) (see chapter about unit root testing) The Johansen cointegration test is carried out by clicking View/cointegration test. In the vector auto-regression estimates window, and the following window will appear:
In this test window we have the option of many different assumptions about the deterministic trend in the model.
Analytics Group
74
The used series may have nonzero means and deterministic trends as well as stochastic trends. Similarly, the cointegrating equations may have intercepts and deterministic trends. The asymptotic distribution of the LR test statistic for cointegration does not have the usual distribution and depends on the assumptions made with respect to deterministic trends. There11 fore, in order to carry out the test, it is needed to make an assumption regarding the trend underlying the data. Summary of the 5 options in the test:
have no deterministic trends and the cointegrating equations do not have intercepts have no deterministic trends and the cointegrating equations have intercepts have linear trends but the cointegrating equations have only intercepts and the cointegrating equations have linear trends have quadratic trends and the cointegrating equations have linear trends:
Note: including extra exogenous variables in the exog variables will lead to wrong critical values. The lag interval specified (for example 1 1) will include one differenced lag , 1 2 will include and .
By selecting 0 in lag and option 2, one should get the following output by clicking OK:
And some more output. So we have the two Johansen test, Trace and maximum Eigenvalue test, and these yield the same results in this example. One can see the sequential test procedure used in the Johansen test. First we are testing the null hypothesis of zero cointegration relationships against the alternative hypothesis of 1 or more cointegration relationships. In the first step we can see that the null hypothesis of none cointegration relationsships is rejected -> next we test at most 1, against at most 2 and so on, until we cannot reject anymore. In the example given above the test finds 2 cointegration relationships in both tests.
11
Analytics Group
75
We could also make a plot of the series to get an idea of whether they share a stochastic or deterministic trend, Click on
open as group/view/graph.
The general vector error correction model with deterministic trend is This can be rewritten into our test equation: ( Where and )
The intuition of this expression is that a change in can come from the time trend, or the error correction part of the expression (the error correction part is the only in parenthesis). The last part of the expression with a summation from up to of lagged values of the differenced dependent variable is used to eliminate serial correlation. Since our dependent variable is differenced in the test equation, we note that: Deterministic trends Quadratic trends The linear combinations have The linear combinations are trending deterministically in in a non-zero equilibrium value
Analytics Group
76
The lag interval in the window above (1 1) corresponds to , this can be seen by looking at the sum operator in our test equation expression: . So if we wanted to test this for p=1, we should set lag interval equal to 0. In the tab cointegration we should specify the correct number of cointegration relationships. We found the number of cointegratio relationships to be two in the Johansen cointegration test in the previous section. By clicking the tab Cointegration we get to include this information in the test and determine which trend should be included.
Analytics Group
77
An explanation for all 5 options can be found under the section Johansen cointegration test. If we want to estimate the model using p=1 (lag interval = 0) and r=2 (meaning two cointegration intercepts) and option 2) in the cointegration tab, we should get the following output:
1.
Represents the long run equilibrium relations. The first cointegration equation, CointEq1, is estimated as which can be rewritten as: The second cointegration equation, CointEq2, is estimated as .
Analytics Group
78
2.
3.
The error correction part represents the short run relations. CointEq1 relates R10Y with R3M. What we can see is that if the value of R10Y lies above its long run equilibrium, then R3M will increase next period. CointEq2 relates R1Y with R3M. Here the short run effect is that if 1Y interest rate lies above its long run equilibrium, the R3M interest rate will fall next period. We notice that the information criteria which we want to minimize is reported as well.
Again we could test for serial-correlation, to see if we have included the right amount of lags (or if we should increase p in our test equation). This is done as before by choosing View/Residual test/Correlogram, if we do that it will be clear to us that we should use at least one more lag in the test ( ).
Clearly we still have a lot of serial correlation, especially in the first period. We will not show it here, but one should repeat the process of estimating the VECM model, but now with .
Analytics Group
79
Where denotes the information set, typically including and its entire history. This sepcification is called an ARCH(1) process. If the conditional variance depends on time, then we say that there are ARCH/GARCH effects. This can be tested with a Breusch-Pagan test for heteroskedasticity (chapter 4 Verbeek 4th edition). Here an auxiliary regression is run on the squared OLS residuals upon lagged squares and a constant and compute T times the . Under the null hypothesis of homoskedasticity ( ) the test statistic is asymptotically Chi-squared distributed with p degrees of freedom.
Now we would like to test whether this process suffers from ARCH/GARCH effects, and if this is the case, how can we best model the variance? 1. First we estimate the above model by OLS. This is done by typing ls Y c Y(-1) in the command field.
Analytics Group
80
2. 3. 4.
Second we must save the OLS residuals from the above estimation. You can use the short cut and type genr e=resid in the command field or click proc/make a residual series. Third we have to compute the squared residuals, which is done by typing genr e2=e^2 and press enter in the command field. The last thing is to run an auxliary regression by the following command ls e2 c e2(-1) .
Analytics Group
81
The
from the auxiliary regression is 0,026709 this should be multiplied with the number of observations.
( )
Now we have established that the conditional variance varies over time. We can now try to estimate the conditional variance by the build in function in EViews. This function will appear by typing ARCH in the command field.
Analytics Group
82
1. 2.
Fist we have to inform EViews which regression we have previously worked with. Type in specification of the estimation Y c Y(-1). Then we have to specify the ARCH or GARCH model. You need to make a number of different ARCH and GARCH estimations and compare the information criterion and statistical significance of the ARCH and GARCH terms in order to evaluate the best fit.
In this example we will only estimate an ARCH(1) and a GARCH(1,1) model and evaluate which of the two has the best fit. You should compare a larger number of ARCH/GARCH specifications when you evaluate the best fit.
Analytics Group
83
Based on the information criterion and the fact that the ARCH effect is insignificant in the ARCH(1) model we will choose a GARCH(1,1) to model the conditional variance. The last thing we are going to do is make a plot of the conditional variance. 1. Click on view and GARCH graph then you can choose between variance and standard deviation.
Analytics Group
84
12 Panel data
Panel data models combine the time dimension and the cross sectional dimension. In panel data we sample the same cross sectional units over time, and thus get a two-dimensional data set. Important advantages of panel data compared with time series or cross-sectional data set is that they allow identification of certain parameters or questions, without the need to make restrictive assumption. For example, panel data make it possible to analyze changes on an individual level. Consider a situation in which the average consumption level rises by 2 % from one year to another. Panel data can identify the result of, an increase of 2 % for all individuals or an increase of 4 % for approximately one-half of the individuals and no change for the other half. In more general terms we try to capture all unobservable time invariant differences across individuals. In this section we will go through four different panel data models, pooled cross section, fixed effect estimation, random effect estimation and first difference estimation. The basic panel data model for cross sectional unit is:
Where is a vector of observable explanatory variables, is the idiosyncratic error with mean 0 (the error varies both over time and individual). is the unobserved time constant characteristics of an individual, and that is the effect we specifically want to control for in our panel data models. A classic example of unobserved characteristics could be an individuals ability (since its non-measurable). ( ) Normally we will use random effects models if the unobserved characteristics are uncorrelated with the ( ) explanatory variables. If we would use either fixed effects estimation or first difference estimation. It is possible for us to use instruments and robust standard errors in panel data models.
The unobserved effects in this model could be worker and managerial ability (nearly impossible to measure). In this job training program we are sampling the job training grants, which were approved on a first-come, first-serve basis, meaning that the unobserved effect might be correlated with .
Analytics Group
85
In this dataset our panel identifier series is the variable (ID value for each cross sectional unit), and our Date series is the year variable. The frequency is set to Annual, because we have yearly data. With the start date set to @first and end date set to @last EViews automatically finds out the start and end date. Alternatively we could manually have written 1987 in start date and 1989 in end date. Clicking OK EViews will create a new variable called dateid, and EViews will now be able to produce panel data model estimates.
Analytics Group
86
1. 2.
3.
We write up our model of interest Normally we just choose to estimate the model by Least Squares, but if we have problems with endogeneity we can choose TSLS, and then a new tab called Instruments will appear where it will be possible to specify instruments. We click on the Panel options tab, and the following window appears:
Analytics Group
87
In the effects specification we can choose between fixed and random effects estimation by choosing Fixed or Random in the Cross-section option. If we choose none we will just get pooled OLS. We start by choosing fixed effect estimation. And click OK - we get an error: near singular matrix.
This is because fixed effect estimation does not accept an explanatory variable that is constant over time, and union is constant over time (mathematically its impossible to inv ert a singular matrix, and that is what causes the error). We fix this by removing union from the regression, and run it again. Output produced:
Analytics Group
88
The constant in the regression assures that the unobserved effect tain the s by clicking View/Cross-section Effects.
The data in the column Effect is our unobserved effect given for each individual in the sample:
The intercept and the unobserved effect are differenced away, since we have lagged the model one period and subtracted it from the original model to obtain the first difference model. As before we have a problem with the explanatory variable, union, therefore we have to exclude it because it is constant over time. We have the choice of including a con-
Analytics Group
89
stant and deleting a dummy, or keeping both dummies but leave out the constant in the estimation. An easy way to take first difference of a variable is writing ( ) in the equation specification:
The effects specification in the tab Panel Options/Cross-section should be set to None. Estimation output with and without constant:
Analytics Group
90
Then we estimate the AR(1) model either by using the Quick/Estimate Equation, or just typing ls du du(-1) in the command window. EViews produces the following output:
The coefficient estimate on DU(-1) should be 0 for the first difference estimator to be efficient, and -0,5 for the fixed effect estimator to be efficient. One should carry out a wald test of DU(-1) being equal to either 0 or -. Overview table: ) Fixed effect est. Consistent; efficient if coef. on DU(-1)=0 First difference est. Consistent; efficient if coef. on DU(-1)=- Normally we choose the estimator that is closest to have fulfilled the efficiency condition. (
Analytics Group
91
The Hausman test can be carried out by EViews after the random effects estimation, by clicking View/Fixed/Random Effects Testing /Correlated Random Effects Hausman Test. The output produced:
The test yields an error because of non-invertible differences in covariances between the two estimators. To be able to carry out the test we need to re-estimate the random effects model without all dummies and time-constant variables like union in our example. The new Random effects specification is: lscrap c grant grant_1. Now its possible to make a valid Hausman test, and it yields the following output:
( ) The null hypothesis of the Hausman test is that both estimators are consistent and thus . We cannot reject the null on a 5% significance level, and therefore we have that the unobserved characteristics and the explanatory variables grant and grant_1 is uncorrelated. Therefore we conclude that the random effects estimator is the most efficient to use in this case.
Analytics Group
92
Where CG is the US annual fourth quarter consumption growth rate (per capita and in real units), and aversion.
The vector of annual excess returns is based on 10 decile portfolios sorted by the book-to-market ratio. The first decile contains the stocks with the smallest book-to- market ratios (growth stocks), while the tenth decile contains the stocks with the largest book-to-market ratios (value stocks). The portfolios are constructed using stocks listed on NYSE, AMEX and NASDAQ. More information is available from Kenneth Frenchs website from which the portfolios have been downloaded. First we have to specify our moment conditions. The population moment conditions are
[ Where
We have 10 moment condition and only 1 parameter to estimate, implying our system is overidentified. In the data set
and
To type the moment conditions into EViews you have to click on object /New object/system and then type in the moments. Then we need to specify which instrument we are using in all our moments. In the system you write @inst c (c is equal to ), then the moments follows. You can set a starting value for c, by typing param c(1) 10 after the moments. In this example the starting value for c(1) is set to 10, which is shown below:
Analytics Group
93
Now our system and specifications are done and we can estimate c(1), Click proc /estimate and choose the estimation method GMM HAC .We also need to tell EViews which weighting matrix we desire (identity matrix or the optimal weighted matrix). In this example we will use identity matrix because we want equal weight of the 10 portfolios.
Analytics Group
94
As you can see the EViews estimates of standard errors are rather high. In the 7 edition of E-Views there are problems when estimating GMM with a system. It seem like EViews cant handle the estimation of standard errors with a n identity matrix. A solution to this will come as soon as possible. The right estimates can be seen below. This estimation is done in th the 5 edition of EViews.
th
The estimated coefficient of c is 87,996 with a s.e. of 43,86. Given asymptotic normality, we are able to reject (marginally), meaning that the coefficient of relative risk aversion is significant.
Analytics Group
95
14 Programming in EViews
EViews is capable of doing various programming tasks. Here we will give a brief introduction to the basic commands, syntax and methodology. EViews help function is very useful and should be used if you are in doubt about any command, and/or function. The helpfunction can be found in EViews: Help/EViews Help Topics. The most efficient way to search for a given command or item, is under Index, where it is possible to type the search query.
Run will open a dialogue box, where you can select additional options. Typically we want to choose Quiet (fast) under the Reporting tab, to enhance the speed of the compiling process. To run the program we simply press the OK button in the
bottom of the dialogue box.
a s q m w d 7 u
annual. semi-annual. quarterly. monthly. weekly. daily (5 day week). daily (7 day week). undated/unstructured.
Analytics Group
96
14.3 Comments
Another useful thing is the possibility of making comments in the code, by using an apostrophe.
wfcreate my_workfile_unstructured u 1 1000 this is a comment, and will be ignored by the compiler
Using comments to comment on your code is best practice. It will ease the reading of the code.
scalar count
We could assign a value to the scalar after the declaration
count = 1
To declare a vector of size n, we type the command vector(n) followed by the name of the vector we want to declare .
vector(n) my_vector
After running this code, we will have an n-sized vector in our workfile. Other similar declarations:
matrix (5,5) my_matrix creating a 5 by 5 matrix rowvector(5) my_rowvector creating a rowvector instead of a columnvector
To fill a value into the vector, we write the name of the vectorobject followed by a dot and fill
vector(5) myvector my_vector.fill 1,2,3,4,5 filling the values 1,2,3,4,5 into the columnvector my _vector
genr price = 0
Analytics Group
97
The generated series will have the size of the chosen workfile size. If we have chosen our workfile as unstructured with 1000 observations, price will be a series of size 1000. If we use a variable in the program that is not intended to be saved in the workfile, we declare it by using an exclamation mark in front of the name of the variable. It could for example be a count variable in a loop (see under loop, next section).
!count = 1 setting the variable count equal to one !count = !count + 1 setting the variable count equal to the previous value plus 1
smpl 1992m1 1993m2 smpl 1 1000 smpl @all reversing the sample size to all observations
equation reg1.ls y c x equation reg2.ls(h) y c x now with white robust standard errors
These equation objects will now be saved in our workfile. It is possible to extract certain information from our equation object. We can for example extract the r-squared or the coefficient estimates from our regression.
Analytics Group
98
@logl @meandep @ncoef @r2 @rbar2 @rlogl @regobs @schwarz @sddep @se @ssr @stderrs(i)
value of the log likelihood function mean of the dependent variable number of estimated coefficients R-squared statistic adjusted R-squared statistic retricted (constant only) log-likelihood. number of observations in regression Schwarz information criterion standard deviation of the dependent variable standard error of the regression sum of squared residuals
Analytics Group
99
cointreg estimate cointegrating equation using FMOLS, CCR, or DOLS. count count data modeling (includes poisson, negative binomial and quasi-maximum likelihood count models). glm estimate a Generalized Linear Model (GLM). gmm estimate an equation using generalized method of moments (GMM). liml estimate an equation using Limited Information Maximum Likelihood and K-class Estimation. logit logit (binary) estimation. ls equation using least squares or nonlinear least squares. ordered ordinal dependent variable models (includes ordered probit, ordered logit, and ordered extreme value models). probit probit (binary) estimation. qreg estimate an equation using quantile regression. stepls estimate an equation using stepwise regression. tsls estimate an equation using two-stage least squares regression. Help for the different estimation methods can be found in Eviews. In TSLS estimation we will have to specify our instruments for the estimation to work.
equation twostep.tsls y c x h @ c j h
The instruments for the regression is given after the @-symbol. We have that c and h and instruments for themselves and x have j as instrument.
14.9 Loops
The idea of making either a for or while-loop is to iterate through the same code, doing a certain task many times. It could for example be that we wanted Eviews to estimate 1000 OLS-regressions in our data set. A for loop that places the values from 51 to 1050 in a vector.
wfcreate my_workfile_unstructured u 1 1000 'creating workfile vector(1000) values 'declaring vector of size 1000 for !i=1 to 1000 'for-loop beginning values(!i) = 50 + !i next
Analytics Group
100
wfcreate my_workfile_unstructured u 1 1000 'creating workfile vector(1000) values 'declaring vector of size 1000 !i = 1 while !i <= 1000 while-loop beginning values(!i) = 50 + !i !i=!i+1 wend
After the while-statement we have a logic expression, and as long as the logic expression is true, we will loop. In the while loop we need to manipulate our controlvariable i ourselves.
Where
'remember to run program in quiet mode (fast) wfcreate Monte_Carlo u 1 1000 'creating workfile vector(1000) values 'declaring vector of size 1000 !m = 1000 'number of simulations for each sample size vector(5) samplesize samplesize.fill 10, 50, 100, 500, 1000 rndseed 1 'seeding the random number generator, making it possible to reproduce results matrix(!m , 5) results 'creating a matrix with dimension 10000 by 5 to store slope coefficients in for !i=1 to 5 'loop for picking out sample size !ss = samplesize(!i) for !j=1 to !m 'data generation genr x = nrnd 'the values of x is drawn from a normaldistribution genr z = nrnd genr e = -x(-1)+z 'generation of our error term according to the specification genr y = 1+x+e smpl 2 1000 'restricting the sample size, because of the lagged value 'estimation equation ols.ls y c x results(!j , !i) = ols.@coefs(2) 'storing the slope coefficient in the j'th row, and the i'th column. next next
Analytics Group
101
( (
) )
Now we can assess the properties of the estimator by double clicking the malpha and mbeta matrix respectively and choose View/Descriptive stats by column.
Analytics Group
102
0= Undecided; 1= Soc.Dem.; 2=Rad.V.; 3=Kons.; 4=SF; 5=CD; 6=Da.Fo; 7=Venstre; 8=Fremskr.; 9=Enhedsl.; 10=Andre Sp04: Your weight Sp05: Your height Sp06: Your mother's height Sp07: Your father's height Sp08: Drinks (Genstande), number of in week 34 Sp09: Average marks (Karakter) at qualifying exam Sp10: Social order 1=Rather important; 2=Important; 3=Very important; 4=Extremely important Sp11: Social justice 1=Rather important; 2=Important; 3=Very important; 4=Extremely important Sp12: Belong to somebody 1=Rather important; 2=Important; 3=Very important; 4=Extremely important Sp13: Selt-realization 1=Rather important; 2=Important; 3=Very important; 4=Extremely important Sp14: Fun and joy in life 1=Rather important; 2=Important; 3=Very important; 4=Extremely important
Age -the age measured in total years Childd dummy variable 0 = no children, 1=one or more children Educatio number years of education completed, including elementary school Exper amount of experience in current job measured in years, Hourwage personal pretax income per hour measured in krones ln_hwage LN(hourwage) that is the natural logarithm on the above described variable
Analytics Group
103
marriedd dummy variable 0 = if not married, 1= if married provinced dummy variable 0 = if the respondent does not work the in province, 1= if respondent work in the province.
Note that we used Excel to modify the existing SPSS file FEMALEPRIVATEWAGE.sav to Eviews format (.wf1). This was done to create the above mentioned dummy variables.
Analytics Group
104
ANALYTICS GROUP
Analytics Group, a division comprised of student instructors under AU IT, primarily offers support to researchers and employees. Our field of competence is varied and covers questionnaire surveys, analyses and processing of collected data etc. AG also offers teaching assistance in a number of analytical resources such as SAS, SPSS and Excel by hosting courses organised by our student assistants. These courses are often an integrated part of the students learning process regarding their specific academic area which ensures the coherence between these courses and the students actual educational requirements. In this respect, AG represents the main support division in matters of analytical software.
Advanced Multimedia Group is a division under AU IT supported by student instructors. Our primary objective is to convey knowledge to relevant user groups through manuals, courses and workshops. Our course activities are mainly focused on MS Office, Adobe CS and CMS. Furthermore we engage in e-learning activities and auditive and visual communication of lectures and classes. AMG handles video assignments based on the recording, editing and distribution of lectures and we carry out a varied range of ad hoc assignments requested by employees. In addition, AMG offers solutions regarding web development and we support students and employees daily use of typo3.
PLEASE ADDRESS QUESTIONS OR COMMENTS REGARDING THE CONTENTS OF THIS GUIDE TO ANALYTICS@ASB.DK